Multi Target Recognition And Position Estimation Of Aircraft Based On Visible Image

With the rapid development of computer vision,target recognition and tracking algorithms based on visible light images have been widely used in video surveillance,human-computer interaction and other fields.This thesis mainly studies the target recognition and tracking algorithms.In view of the shortcomings of the traditional algorithms,this thesis proposes an improvement scheme to achieve multi-target recognition of aerial images and multi-target tracking of aerial videos.Based on the aerial images of the aircraft,this thesis theoretically researches the method of estimating the position of the ground target.Firstly,this thesis studies the mean filtering algorithm,median filtering algorithm,and Gaussian filtering algorithm to remove image noise.Through simulation experiments,the denoising effects and characteristics of the three filtering algorithms are obtained.Considering the impact of haze on the image pixel quality,this thesis researches the image defogging algorithm of dark channels prior,and defogging simulations are performed on haze images and videos.Aiming at the deficiencies of the traditional video defogging algorithm,an improved video defogging algorithm is proposed and verified by simulation.Then,in terms of multi-target recognition,firstly,an improved feature extraction algorithm of multi-target recognition technology,was proposed in the traditional machine vision framework,and it was applied to the traditional machine vision multi-target recognition algorithm.The multi-target recognition simulation is carried out for the ground head up image and the low altitude aerial image respectively.The simulation results show that the accuracy of recognition needs to be further improved.Next,considering the limitations of traditional machine vision,this thesis studies a multi-object recognition algorithm based on neural networks.Through the analysis and simulation of the algorithm of YOLO and SSD,the results show that YOLO and SSD algorithms have poor anti-occlusion capabilities,missed detection and recognition errors.Therefore,an improved neural network algorithm is proposed,and the simulation results are compared with other multi-target recognition algorithms in this thesis,which effectively improves the precision and accuracy of the multi-target recognition algorithm.Finally,this thesis focuses on the KCF target tracking algorithm.Because the KCF algorithm's target frame cannot be adaptively scaled and has poor anti-occlusion ability,this thesis designs a multi-scale module and an anti-occlusion module,and proposes an improved multi-target tracking algorithm to achieve stable multi-target tracking of aerial video.Aiming at the problem of image target position estimation,this thesis studies the camera imaging principle and coordinate conversion,and through the image ranging method based on similar triangles,realizes the function of target positioning and ranging through two-dimensional images.